A chat transcript is not enough for work that runs over hours or days. Here is what we keep instead, and why it makes agents easier to pause, resume and check.
A thought to take with youKeep a small, structured record of where the work stands. Let the conversation be a log, not the memory.
What goes wrong with transcripts
The simplest way to give an agent memory is to send it the whole conversation every time. That works for a short chat. It starts to fail when a task runs across many steps, tools and days: the transcript grows, older details get buried, and the agent spends attention re-reading things that no longer matter.
It is also hard for a person to answer a basic question from a transcript: where is this job right now, and what is it waiting for?
What we keep instead
We keep a small structured record next to the conversation: the goal, the steps done, the step in progress, what each tool returned, and any open approvals. The agent reads that record first and only pulls in older messages when it needs them.
Each finished step is saved as a checkpoint. If a process stops, the work resumes from the last checkpoint instead of starting again.
Why it helps the people around it
A clear record of state is something a supervisor can read in seconds. It makes it obvious what the agent did, what it is about to do, and where a person needs to step in.